A Google-Like Approach to Mammograms

A new approach to mammograms speeds up computer automated "second opinion" interpretations of breast images. Just as a Google search first returns a list of only those websites that it determines to have the most important and useful information on the words entered in the search, so medical researchers want to speed up the computer search for suspicious-looking breast masses. Georgia Tourassi (Georgia.tourassi@duke.edu) and her colleagues at Duke University employ such a Google-like approach that can retrieve useful information from steadily growing mammogram-image databases more rapidly than before.

Increasingly being used in clinical settings, knowledge-based computer-assisted detection (KB-CAD) systems compare a mammogram image to those of known cases of breast cancer in order to aid radiologists in their diagnosis. When a new, unknown case is presented for analysis, the KB-CAD system compares the case to mammography images in the database. If the unknown case is visually similar enough to a known case of breast cancer, then this would suggest the presence of cancer.

Traditionally, KC-CAD systems compare the mammogram image under investigation to every image of breast cancer in a computer database. Although diagnostically accurate, this practice becomes inefficient as image databases increase in size. A larger wealth of images provides more information from which the systems can draw upon for analysis, but comparing the mammogram in question to every image becomes inefficient. Therefore, the Duke researchers incorporate an additional, "Google," approach. They compare the mammogram only to selected images that are most highly ranked for their information content.

The selection of the most informative mammogram cases is performed using a strategy based on the concept of "image entropy." Image entropy represents a measure of the disorder or complexity in the image. An image that is all black or white has zero entropy. An image of a checkerboard has low entropy--it consists of an equal number of light and dark pixels. Complex images with more uniform distributions of many pixel intensity levels have higher entropy and are considered more informative in the context of the Duke system. However, normal breast tissue can be as complex as a tumor. According to Tourassi, who reported her results at this week's meeting of the American Association of Physicists in Medicine in Orlando, this is precisely the reason mammographic diagnosis is such a challenging task. The Duke approach includes normal cases as well in the decision-making process.

Applied to an existing database of 2,300 mammography images, the Duke method compared a sample mammogram to 600 images it ranked as most informative. This cut down the time the CAD system took to analyze the mammogram by one-fourth, to less than 3 seconds per query. In the next year, the researchers expect to follow up their pilot study with a larger investigation to evaluate the clinical impact of this new approach.

Meeting Paper: TU-D-330A-8
AAPM meeting virtual pressroom
Contact Georgia Tourassi, Duke University
georgia.tourassi@duke.edu